You Reek of AI

Six months ago, someone receiving the email above wouldn't have looked twice. Nothing "off" about it would have registered. But today, things are different. Even an avid user of ChatGPT or Claude could probably not put their finger on exactly what gives it away, but they just "know it when they see it."

Alongside security vulnerabilities and weak governance, this is one of the biggest threats facing businesses that adopt AI without proper guardrails: it leaks. Not your data but your voice. And like a bad cheese, once people can smell it, they can't un-smell it. Let's go through it line by line.

Why AI Leaks Your Voice

"I wanted to follow up on my previous email — just checking in."

AI is addicted to the em-dash, a character most people don't even know exists. It's not on your keyboard (its one of those alt+[code] keys), and while applications like Microsoft Word will sometimes auto-correct a hyphen to one, AI deploys them almost aggressively (and it can get downright hostile about them). You'll also notice the hedging that starts here: "just to check in" softens what could simply be "following up on my email from Tuesday." In a real business email, the subject line alone would do more, something like "Mark re: Cooper Automotive" tells you everything. AI tends toward the generic.

"No rush at all, but I did want to briefly reiterate..."

This is qualifier stacking. The AI has been 'trained' to avoid sounding pushy, so it layers one softener on top of the next until the sentence buckles under its own politeness. A human writer might hedge once. AI hedges until it sounds like it's apologizing on your behalf for merely existing.

"e.g. around next steps or potential timelines, if helpful"

This one is subtle but revealing. AI frequently misuses "e.g." and "i.e.", often interchangeably, because as a statistical pattern matcher rather than a cognitive reasoning engine, it has simply seen patterns like this in its training data and reproduces them without "understanding" the distinction. The phrase "if helpful" at the end is another tell: a throwaway that adds nothing and exists only to avoid commitment language.

"If it makes sense, we could also connect for a quick call. Otherwise, I'm happy to continue the conversation over email."

This has become one of the most recognizable AI business cliches of the moment. It's assertively non-committal. A human would say "would love to jump on a call this week" or even just "happy to set something up." But more telling is the second half: why offer email as an alternative to a call? What other options were on the table? AI covers all its bases because it doesn't actually know (or care) what you want. It hedges from its core.

The Totality Is the Problem

Now, you might be thinking: I've written emails with phrases like these. Of course you have, we all have. The problem isn't any single element in isolation. It's the totality. Together, these patterns form a statistical signature that anyone who spends real time working with large language models has started to subconsciously recognize. It's called AI slop, and it's quietly eroding trust in business communications that should feel personal.

How to Fix It

Prompt engineering is a weak guardrail. LLMs are probabilistic pattern-matching machines, and no matter how carefully you command them, you cannot deterministically control their output. Treat them accordingly.

The practical implication: enforce rules of engagement in code, not in prompts. Any endpoint that can push an email, publish a LinkedIn post, or trigger outbound communication of any kind should sit behind a human review gate, not (read: never) a system prompt. Deterministic logic protects your brand. Stochastic text generation can destroy it.

The stakes are highest with people who already know you.

When an AI generated message lands in the inbox of a colleague, client, or investor, they aren't reading it in a vacuum. They're reading it against years of accumulated knowledge about how you write, what you emphasize, and how you sound. The reputational cost of sounding like a bot to someone who knows you is far higher than whatever efficiency you gained by skipping the review.

Prompt with a goal, not just a task.

The single biggest driver of AI slop is vague prompting. When you hand a model a task without context, it fills the gaps with the most statistically average "version" of what you asked for. That's how you get emails that sound like they were written by Rose the chatbot at 2am.

The fix is simpler than most people think. Before you write a single word of your prompt, get clear on what you actually want to happen as a result of this message. Not what you want the email to say, but what you want the recipient to do, feel, or understand after reading it. That distinction is everything. A model given a task writes toward completion. A model given a goal writes toward an outcome.

The Four Qualifiers Every Prompt Needs

Build these four qualifiers into every prompt template you use:

• GOAL: What outcome do you want from this message?
• CONTEXT: Who is this person and what's the relationship history?
• TONE: How should this land? Urgent, casual, formal, warm?
• CONSTRAINT: What should it not say or do?

Mary Can Smell It

Mary can smell it. So can your best clients, your investors, and anyone else who's spent real time anywhere lately. The good news is that the bar for not reeking of AI is lower than most people assume. It doesn't require abandoning these tools in totality. It just requires using them with enough intention that the output sounds like it came from a person who actually gave a damn.